TRIDENT: Efficient Triple-Task Learning of Dehazing, Depth, and Uncertainty Estimation for Underwater 3D Robot Visual Perception

Underwater visual systems greatly suffer from blurry textures and low color contrast due to the inevitable light propagation. These issues can significantly degrade the perception for stable robotic operations. In this paper, we introduce a novel learning-based sensing system that tackles the multidimensional vision tasks underwater; concretely, we deal with image enhancement, depth estimation, and uncertainty for 3D visual systems. Also, we propose a TRIDENT model that takes three objectives with fast and light weight; TRIDENT consists of parallelized three decoders and one backbone structure for efficient feature sharing. In addition, it is designed to be trained on a three-stage hierarchical feature space to express complex parameterization. In experimental evaluation on several standard datasets, we demonstrate that TRIDENT significantly outperforms other existing methods on image enhancement and depth estimation. It also has better efficiency than the others for both memory size and reference time, despite performing three tasks. Finally, we prove that our proposed joint learning is robust to feature matching and has no limitations in extending 2D to 3D vision tasks.

We tried to form a synthetic dataset with depth D ∈ [1.5, 3, 5, 7.5, 10, 15] according to the water types in 'Inherent optical properties of Jerlov water types' paper. For water types I, IA, and IB, the depth was [3, 5, 7.5, 10, 15]. For type II, the depth was [3, 5, 7.5, 10]. For III and IC types, the depth was [3, 5, 7.5]. For 3C and 5C, the depth was [3, 5]. Finally, for 9C, it was [1.5, 3]. The ground dataset used was a dataset with an object detail called 'diml'. A total of 48,000 synthetic data were generated, of which 47,000 were selected for the training set and 1,000 for the test set.

This dataset was created through distortion optical modeling, and has Enhanced Images and Dense Depth Images as ground truth. This dataset will be of great help to those working on Underwater Depth Estimation and Underwater Image Enhancement.

Data in the Joint-ID_Dataset

• Synthetic Underwater Train Set: 47,000
• Synthetic Underwater Test Set: 1,000

The structure in the Joint-ID_Dataset

/ train
    /    LR
    /    ...
         / color
             / <enhanced images>.png
         / depth_filed    #
             / <dense depth images>.png
    /    synthetic
         / <synthetic distorted images>.jpg
/ test